Research map / Label-efficient and robust learning / Recognition and long-tail

Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes OWL-GPS, a relevance-guided active search and online meta-learning framework with a concept-conditioned CVAE for budget-constrained PFAS contamination discovery.

Problem

Discovering under-observed target regions (e.g., PFAS contamination hotspots) under tight sampling budgets, with sparse labels, non-stationary distributions, and no revisitation or replay of past samples.

Approach

Uses a CVAE-based concept-conditioned relevance vector (over Gram-Schmidt-orthogonalized domain concepts) to jointly guide uncertainty-based sampling, relevance-aware meta-batch formation with core and reservoir buffers, and online meta-updates.

Key results

• On the 2019 PFAS data, the full framework reaches 98% SR and 65% F-score at budget C=50, versus 94% SR for the best baseline (UCB/Prithvi).
• On the 2022 transfer setting, the method reports 86% SR and 54% F-score, above OML (84% SR, 51% F-score).
• Removing the relevance encoder drops F-score from 65% to 56% on the 2019 data.

Surprising findings

• Random meta-batch selection and removing relevance-guided sampling both substantially reduce SR, indicating the relevance-based components drive most gains.
• On the 2022 transfer setting, the proposed method's accuracy (75%) is not higher than UCB (75%) and GA/AL have comparable accuracy, so gains are mainly in SR and F-score.

Limitations

• Relies on domain-specific concept drivers, which may limit use in unstructured tasks.
• Evaluation is on PFAS and a sparsified land-cover task; no public code or data release is stated.
• Several theoretical propositions rely on strong assumptions (strong convexity, unbiased gradients) that may not hold for neural policies.

Experimental setup

Datasets: 2019 PFAS dataset 2022 PFAS data land cover data

Benchmarks: 2019 PFAS dataset 2022 PFAS data sparsified land cover (water class)

Metrics: Success Rate accuracy F-score precision recall

Baselines: GA DIG Prithvi UCB AL AML OML

Builds on: model-agnostic meta-learning conditional variational autoencoder focal loss Prithvi

Closest papers by meaning

• Active Target Discovery under Uninformative Priors: The Power of Permanent and Transient Memory (NeurIPS 2025 · similarity 0.85)
• Online Feedback Efficient Active Target Discovery in Partially Observable Environments (NeurIPS 2025 · similarity 0.83)
• FireScope: Wildfire Risk Raster Prediction With a Chain-of-Thought Oracle (CVPR 2026 · similarity 0.81)
• Environment-Robust Representation Learning with Empirical Bayes (NeurIPS 2026 · similarity 0.81)
• Decision-aware Training of Spatiotemporal Forecasting Models to Select a Top-K Subset of Sites for Intervention (ICML 2025 · similarity 0.81)
• GST-UNet: A Neural Framework for Spatiotemporal Causal Inference with Time-Varying Confounding (NeurIPS 2025 · similarity 0.81)